Uber caps retry storms by making services claim ownership of errors

Sources: primary, discussion
Summary: Uber states that retry budgets alone still let a deep call chain amplify one failure, because every hop retries against its own budget. The post describes marking which service owns an error so callers stop retrying failures they do not own, and reports the maximum retry storm radius falling from 25 hops to 3.
Why it matters: The change is in what an error carries rather than in added capacity, so it applies to any deep RPC chain with per-hop retries.